PK Wiki

source · sources/Making AI Work — Leadership, Lab, and Crowd.md

Making AI Work: Leadership, Lab, and Crowd

status: active · created 2026-06-02 · 1 source · #ai-and-work #organizational-change #adoption #ai-and-work #mollick

Ethan Mollick's One Useful Thing essay (2025-05-22) arguing that AI's individual productivity gains don't become organizational gains without deliberate organizational innovation — coordinated across three roles: Leadership, Lab, and Crowd.

Summary

Mollick opens with four facts he says explain AI adoption:

  1. AI boosts work performance — self-reports (a Denmark knowledge-worker study: AI halved time on 41% of tasks; a US survey: workers said AI tripled productivity, 90-min tasks → 30 min) backed by controlled experiments across product development, sales, consulting, coding, law students, and call-center work.
  2. Lots of people already use AI at work — Denmark: 65% of marketers, 64% of journalists, 30% of lawyers; US: 30% (Dec 2024) → 40% (Apr 2025); ChatGPT is the 4th most-visited website.
  3. More transformational gains are available than people realize — deep-research reports, emerging agents, increasingly capable systems.
  4. Companies are not capturing these gains — only small/moderate firm-level gains, and no major impact on wages or hours worked as of end-2024 (NBER).

The reconciliation (the core argument): individual performance gains do not naturally become organizational performance. Org gains require organizational innovation — rethinking incentives, processes, and the nature of work — but companies' "muscles for organizational innovation have atrophied" after decades of outsourcing it to consultants and enterprise-software vendors. No one has a playbook; even AI labs "release models without knowing how they can be best used." "We are all figuring this out together" — so advantage goes to whoever figures it out fastest, by harnessing Leadership, Lab, and Crowd.

Leadership — set urgency and a vivid, concrete picture of the AI future (citing colleague Andrew Carton's research that workers change for vivid images of the future, not bottom-line statements). Anticipate task-level change ("AI is not currently a replacement for most human jobs, it does replace specific tasks"), and rethink why tasks are done at all once their cost collapses.

The Crowd — employees who discover uses through trial and error; experienced workers judge AI usefulness in ways outsiders and AI-savvy juniors can't. But official tool use stalls at ~20% while 40%+ use AI privately → the Secret Cyborgs problem (hidden use, for rational reasons). A Leadership fix: safe experimentation zones, incentives, no-layoff reassurance, leaders modeling use.

The Lab — a centralized, ambidextrous team (explore + exploit) staffed from the Crowd, focused on building: distribute Crowd prompts fast; build your own organizational benchmarks (even "vibe benchmarks" — e.g. testing the Manus agent on a fictional startup analysis); "build stuff that doesn't work yet" to be ready when models cross thresholds; and build provocations that jolt people into seeing the potential.

Re-examining the organization — orgs were built around scarce human intelligence; on-demand intelligence shifts the bottleneck (The Bottleneck Shift): "the bottleneck isn't the research anymore, it's figuring out what research to do… the limitation isn't programming speed, it's understanding what to build." Treat AI adoption as an organizational learning challenge, build feedback loops between the three roles, and start now "while everything is still messy."

Key claims

  • Individual ≠ organizational gains. "AI use that boosts individual performance does not naturally translate to improving organizational performance." (Individual vs Organizational AI Gains)
  • No playbook exists. Even AI labs "release models without knowing how they can be best used"; consultants/vendors can't supply a generalized answer "at least for a while."
  • Secret Cyborgs. Official AI use maxes ~20% while 40%+ use it privately; workers hide use because the incentives to stay quiet outnumber the incentives to disclose. (Secret Cyborgs)
  • Tasks, not jobs. "AI is not currently a replacement for most human jobs, it does replace specific tasks within those jobs."
  • Build your own benchmarks. Official benchmarks (trivia/math/coding) don't tell you which model is best at your tasks; develop org-specific ones, "vibes alone" acceptable.
  • The bottleneck shifts to judgment. When execution is cheap, the scarce input becomes deciding what to do. (The Bottleneck Shift)

Convergence with The AI Daily Brief — Why Agents Still Need Humans: Mollick's "bottleneck shift" and Dan Shipper's "humans know what needs to be done right now" (AI Commoditizes Expertise) are the same observation reached independently — strong mutual reinforcement, not contradiction.

Connections

Source

The wiki is the source. This door reads it at build time and never writes to it.